Interview

Why NVIDIA gives away its best models

Kari Briski

VP of Generative AI Software, NVIDIA

21 minHosted by Kim IsenbergRecorded at NVIDIA GTC 2026 in San Jose

Watch on YouTube

About this conversation

At GTC 2026 in San Jose, Kim Isenberg talks with Kari Briski, NVIDIA’s VP of Generative AI Software, about Nemotron 3 Super — NVIDIA’s 120B-parameter open model that activates only 12B parameters at inference.

Kari explains the hybrid Mamba-Transformer architecture, why NVIDIA trained in FP4 precision from day one, how speculative decoding delivers speed gains, and why NVIDIA releases full training recipes, datasets and environments to the open-source community — and why its business model makes that openness possible.

Topics

  • Nemotron 3 Super: mixture-of-experts architecture explained
  • Hybrid Mamba + Transformer: going beyond pure transformers
  • NVFP4 training from scratch: risks and results
  • Multi-token prediction and speculative decoding in practice
  • Configurable reasoning budgets
  • Open-source strategy: recipes, data and trust through transparency
  • The Nemotron Coalition with Mistral AI

Chapters

  1. 0:00Introducing Neatron Super
  2. 1:47Architecture and efficiency
  3. 3:38Optimization and precision
  4. 5:53Decoding and reasoning
  5. 7:18Post-training workflows
  6. 8:57Reinforcement learning
  7. 11:17Open source philosophy
  8. 14:46Safety and auditing
  9. 16:38Reflecting on GTC announcements
  10. 19:35The Neatron coalition

Kim Isenberg is co-founder and Editor-in-Chief of Superintelligence, an AI newsletter published every weekday.

Read the Newsletter

More interviews

All interviews